I Made a Viral AI Love Story (500M Views) — Steal My Prompts

Overview

The video presents an end-to-end AI “love story” production workflow (described as going viral, with a claimed ~500M views). It uses multiple AI tools—most notably:

• Claw / Claude for prompt/script generation
• Cinema Studio for image/video generation, including references to model styles such as “Cedence / Cance 2.5”
• Additional tools referenced for sketches and location generation (e.g., GPT Image / GPT-image-style, Soul Cinema)

The creator claims this workflow avoids common “generic sloppy AI” problems and can be used to create a paid promo-style film.

---

Main Tutorial: Full AI Love-Story Pipeline (Prompts + Order)

Goal

Create a polished, cinematic AI love story using an ordered workflow:

1. Assets
2. Locations
3. Scenes
4. Render
5. Edit

Tools mentioned in the workflow

• Claude (and a “script/skill” reference) to turn scene descriptions into detailed generation prompts
• Claw to create structured prompts / convert character sheets into height/size guidance
• Cinema Studio for generating face/character sheets and running video batches
• Other referenced tools:
  • GPT Image / GPT-image-style tools for sketches
  • Soul Cinema for location generation
  • Multiple character-sheet model options (e.g., Cream 5 Pro, GBT Image 2, Nana Banana)

---

Key Technical Problems + Fixes (Methods)

1) Character sheets: avoid “lookalike actor” face errors

Common failing approach

• Using photo references directly to generate a face/character sheet can produce an unintended generic/actor-like face.
• It may also cause weird face texture issues.

Fix: hybrid head-swap method

• Generate the character sheet using the pipeline for body/costume/silhouette (preserve these).
• Erase the generated head from the main character render.
• Paste the real face photo onto the head in the portrait panel.

Result

• More natural faces across consistent outfits and takes.

---

2) Locations: location choice strongly determines realism (~70% claim)

Key claim

• About 70% of final video quality comes from the location.

Location generation process

• Use Soul Cinema for location generation (variety).
• Generate multiple batches.
• Select the batch that matches constraints such as:
  • space for crowds
  • room for action
  • lighting tone compatibility

Example criteria used to pick the “right” location

• Must allow actions like:
  • packed crowd
  • shoulder bump
  • long sprint
  • leap onto a moving ship
• Avoid:
  • overly cluttered layouts where background details “turn into mush”
  • wrong lighting color casts that would “leak” into later scenes

---

3) Scene prompts: “physics-aware” prompting for correct interactions

Common prompting mistake

• The model may misread the beat (e.g., turning a shoulder bump into a full hug / too-close contact).

Two fixes

1. Explicitly describe physical mechanics
   • momentum
   • feet planting
   • suitcase swing / collisions and exact collision motion
2. Force motion for background/NPCs every frame
   • avoid crowds staring into space

---

4) Sizing / spatial consistency (“giant vs hero” scale problem)

Problem

• Scale inconsistency across generations:
  • “giant” changes size each batch (sometimes much taller/shorter).

Fix: create a height-constraint reference

• Use Claw to convert the character sheet into a structured prompt aligning:
  • hero’s head with the giant’s mid-thigh
• Use GPT Image to generate a simple pencil sketch with two outlines (hero vs giant) showing relative height.
• Feed the sketch to Claude, then regenerate.

Result

• Better size consistency and “spatial logic.”

---

5) Complex space in scenes: draw storyboards instead of relying only on text

Problem

• Large multi-shot action sequences (e.g., train robbery / Western chaos on horseback) often had spatial/logic errors.

Fix: “If it’s about spatial logic, stop describing it and just draw it.”

• The creator sketches storyboard-like frames per shot.
• Store sketches as elements and attach them to the prompt conversation.
• Claude reads geometry from drawings and story from text, producing prompts that match the drawn layout.

Result

• Geometry stays consistent across batches (e.g., horse remains off rails; diagonal jumps land correctly).
• Best frames can be stitched in editing.

---

6) Split long scenes into multiple prompts

Claim

• Reliable “perfect 30-second renders” don’t exist consistently.
• Render in pieces and stitch in the edit.

Example splitting cases

• Giant fight + acting
  • one prompt for fight physics
  • another prompt for acting beats (faces/expressions)
• Carnival/masquerade moment
  • split into three 15-second prompts to give interactions and attention changes “room to breathe”

Reasoning

• One prompt can’t optimize both complex action and acting/face performance at the same time.

---

7) Model selection for character sheet quality

Test described

• Run the same character-sheet prompt through multiple models and compare outputs.

Takeaways (examples)

• For one character: Cream 5 Pro gave better costume texture/wear consistency.
• For the other: GPT Image 2 produced better, more consistent curls across angles.

---

8) Background crowds: “casting extras” to prevent mush/randomness

Problem

• Pirate-battle backgrounds became random “mush,” with inconsistent faces and sometimes barely human figures.

Fix

• Treat background generation like real filmmaking casting:
  • Generate 10 distinct pirates
  • Pack them into a single “crew” element so each deck shot reuses the same characters/assets

Result

• Background characters remain consistent and the scene looks more coherent.

---

Audio Conditioning for Performance Accuracy (Final-Scene “Elevator”)

Key asset: audio/music reference

The creator emphasizes a sound reference that may not show on screen.

Problem without audio attachment

• Without attaching audio, the model may hallucinate a new melody each take.

Fix

• Have a friend record a short voice memo humming the main theme.
• Upload that audio reference into the prompt so the character matches the exact notes.

Demonstration included

• Without audio reference: incorrect rhythm / out-of-tune results.
• With audio reference: correct execution.

Claim

• Works with any track: upload your preferred song as reference and apply the same method.

---

Product / Feature Angle

The creator promotes a custom app that:

• takes user photos (including the bride’s photo),
• generates an AI love story using the described pipeline,
• and is positioned as producing an “exact same love story built for you” workflow.

---

Reviews / Guides / Tutorials Emphasized

The content is framed as a full step-by-step workflow guide covering:

• the pipeline: assets → locations → scenes → prompt generation → batching → editing
• “how to fix” instructions for failure modes, including:
  • sloppy AI faces (head swap)
  • incorrect physics/interaction beats (mechanics prompts + moving NPCs)
  • scaling consistency (sketch-based height references)
  • spatial action logic (draw storyboards)
  • overly long prompts (split renders)
  • crowd consistency (cast-and-pack crew element)
  • audio-driven performance (audio reference conditioning)

---

Main Speakers / Sources (As Stated)

• Speaker: Adil
• Primary tools mentioned: Claude, Claw, Cinema Studio, Soul Cinema
• Additional named models/assumed tools:
  • Cream 5 Pro, GPT Image 2, Nana Banana, and a GPT image sketch tool